Give administrators better work

I want administrators to spend less time moving information between systems and more time making sure the work is done properly. That’s the thinking behind Zero Nine.

I started in administration before moving into analytics engineering and software engineering. I know how much effort sits between receiving a request and completing it. Find the record. Check what’s missing. Ask someone for it. Update another system. Follow up again.

Each step is small. Together, they become the job.

AI gives us a way to hand more of that routine work to software. The opportunity I care about is what the administrator gets to do next.

Give the routine work a clear set of rules

Take supplier onboarding as an example.

A request arrives by email. Someone checks the documents, looks for missing fields, creates a record and sends it for approval. Much of that work follows an established process. A system can collect the information, check it against requirements and prepare the next step.

But a duplicate supplier, conflicting details or an unusual request needs a different response. The system should stop at the boundary of its authority and give a person enough context to decide what happens next.

That boundary is part of the design. So is the evidence behind the recommendation. An escalation that says “please review” and leaves someone to reconstruct the whole request hasn’t saved much work.

flowchart TD
  A[Request arrives] --> B[Check records]
  B --> C{Authorised?}
  C -->|Yes| D[Run routine task]
  C -->|No or uncertain| E[Human review]
  E --> F[Resolve exception]
  D --> G[Record outcome]
  F --> G
  G --> H[Sample routine work]

Put administrators in charge of quality

The people doing the admin often know where a process breaks. They know which fields are unreliable, which requests need a second look and which exceptions keep coming back.

That knowledge should shape the system.

As routine steps become automated, I see the role shifting towards quality assurance and escalation. Administrators can set acceptance criteria, review unusual cases and investigate recurring failures. They can improve the process instead of spending the day keeping it moving by hand.

Review needs its own design. If someone has to reread every document and repeat every check, we’ve added an approval queue to the same workload. Routine cases need checks we can rely on. Exceptions need a clear reason for escalation, the relevant source material and a person with authority to resolve them.

Sampling completed work matters too. A quiet exception queue doesn’t prove the system is getting things right.

Take the load off the person reviewing

A system can finish the task and still leave the administrator mentally overloaded. If they have to remember which requests are waiting, compare five tabs and reconstruct why an agent acted, the work has only changed shape.

Cognitive offloading describes using actions and tools outside our heads to reduce the mental work a task requires. For business software, I want the system to remember the state of the work, keep track of commitments and bring the relevant evidence to the person making a decision.

Show what needs attention, why it was flagged and what the person can do next. Keep the source records and action history one step away. A summary earns its place by reducing the effort of checking it.

For that supplier request, the administrator should see the conflicting details and the approval needed. They should not need to understand how the agent retrieved each document before deciding what to do. They must still be able to inspect those documents, challenge the recommendation and stop the process.

That’s the balance I want: less to keep in your head, with enough visibility to remain in control.

The data work still has to happen

An agent needs access to the right records, a way to act in business systems and rules about what it is allowed to change. It also needs a record of what it did, so someone can investigate a mistake and correct it.

This is where my experience comes together. Administration gives me an understanding of the job. Data engineering gets the information into a usable form. Software engineering connects the decisions to actions.

A good model helps. The surrounding system determines whether the work gets completed reliably.

At Zero Nine, we start with the process and take responsibility for the result, including the handoffs and exceptions. Sometimes that calls for an AI agent. Sometimes a form, a data connection or a straightforward rule solves the problem.

How we’re making this real

Paul is the product we’re building around it. Paul runs recurring administration workflows across the tools teams already use. People approve decisions, and actions stay on the record. The aim is to give the team a clear view of the work and a manageable set of decisions to make.

Measure the work people get back

I would judge an implementation by the time a request takes, the rework it creates and how much attention it still demands from the team. Counting agent runs tells us very little on its own.

The model below makes the review cost visible. Change the proportion of requests escalated and the time each takes. A system that sends too much work back can consume more human time than the manual process.

The change also needs support. People need time to learn the system, permission to challenge it and authority to resolve the cases it sends them. Moving someone into a quality role means giving them control over quality.

That’s the ambition for Zero Nine: systems that take on routine administration, with experienced people responsible for the standards and decisions. Start with one repetitive process. Make its rules explicit. Give its exceptions a proper owner. Then measure whether the team has actually got time back.